Mamba-SAC: selective state space modelling for robot force-guided trajectory tracking
Chunhu Bian*, Yuxuan Guo*, Wenxuan Zhang, Yujie Zhang, Jinyue Liu, Xiaohui JiaPurpose
Force-guided trajectory tracking is partially observable because safe contact depends on recent force/motion history, not only on the current sensor sample. This study aims to evaluate efficient sequence modelling for real-time robot force control.
Design/methodology/approach
The authors integrate the Mamba selective state-space encoder with soft actor–critic (SAC) and a 100-Hz Cartesian impedance control interface. The evaluation compares SAC, long short-term memory-SAC, Transformer-SAC and temporal convolutional network-SAC on a 7-degrees of freedom force-guided robot, with nominal tracking, physical metrics, latency benchmarks and off-nominal stress tests.
Findings
Across five seeds, Mamba-SAC achieves the most consistent nominal reward (−222.5 ± 3.1) and the lowest learning-based steady-state force variation (3.50 ± 0.09 N), whereas Transformer-SAC gives a slightly lower mean trajectory root mean square error. Mamba-SAC also has the fewest sequence-baseline parameters (0.246 M) and near-constant graphics processing unit latency (0.370 ms at L = 32). Additional trajectory, stiffness, disturbance and filtering tests show more stable regulation within the nominal-to-mildly perturbed envelope, while safety filtering reduces extreme errors during high-risk workspace-exit episodes.
Originality/value
This paper integrates selective state-space history encoding into the force-guided SAC control stack, provides a control-oriented interpretation of the learned Δt as a content-adaptive history filter and systematically characterises its operating envelope through real-robot experiments.